Require human review for any inbox reply that involves a complaint, money, safety, a legal term or a public argument. Let automation send only repeat questions that have one right answer.
TLDR: Sort your inbox into four risk tiers. Automate tier 1, let AI draft and a human send tier 2 and 3, and keep tier 4 fully human. Spot check a sample every week.
Four risk tiers for inbox replies
| Tier | Example message | Who sends |
|---|---|---|
| 1. Repeat question | "What are your opening hours?" | Automation, from an approved template |
| 2. Simple request | "Can you send the link to the guide?" | Automation drafts, a person approves in bulk |
| 3. Sensitive | "My order is late and I am annoyed." | A person, with an AI draft as a starting point |
| 4. High stakes | Refund demands, legal threats, safety claims, press | A person only, escalated to a lead |
Start by sorting your last 100 messages into these tiers. Most teams find that about half are tier 1, which is where automation earns its place.
Six triggers that stop automation and call a person
Set these as hard stops, not suggestions.
- Words of anger or urgency, such as "unacceptable" or "scam".
- A refund, cancellation or compensation request.
- A mention of a lawyer, press, regulator or injury.
- The same person writing three times in one day.
- A message in a language your template does not cover.
- Any reply that would quote a price, date or promise not in the template.
When one trigger fires, the thread goes to a named person and the automatic reply stops. Our post on the handoff-to-inbox matrix shows how to route those threads to the right owner.
Where automation is safe on social
Comment automation works best when the reply is the same every time. For example, a viewer comments a keyword on a post and gets a link by message. In Mydrop, trigger-word automations handle this kind of job, and the social inbox keeps every other thread in one place for your team.
Write the template once, have it approved, and then the reply is as safe as the template. That is the whole point of tier 1.
Spot check a sample every week, then adjust the tiers
Review does not end at go-live. Every week, open 20 sent automated replies and ask three questions: Was the answer right? Was the tone right? Should this have been escalated?
Move any message type that fails twice up one tier. Move a type down only after four clean weeks. Log the change so the next person knows why.
For the wider workflow around inbox assignment, see our guide to automating social inbox management. For the same rule applied to posts, read when to force human review on automated posts.
Write the escalation rule so the whole team can follow it
Put one page in your handbook with three parts: the four tiers, the six stop triggers and a named owner for each tier. Add a rule for out-of-hours: tier 3 and 4 messages wait for a person, and tier 1 can still be answered.
Add a response-time promise as well, for example "sensitive threads get a human reply within 4 working hours". A rule with a clock on it is far more likely to be followed than a vague one.
Review the page every quarter, using the spot-check results as evidence.
Sort 100 messages into tiers before you switch anything on
Export or scroll back through your last 100 inbox messages and label each one 1 to 4. Automate only the tier 1 types, write the six stop triggers into your handbook, and book a 15-minute spot check for next Friday.
Google review
Trustpilot review
Product Hunt comment
